1st IEEE Workshop on Performance Evaluation of Tracking and Surveillance (PETS2000)

This paper presents the first results obtained from the statistical object tracker developed at Duisburg University for the test data sequence provided by the organizers of the PETS2000 workshop. Our system differs from many other systems in that respect that it dos not use any motion information at all for calculating the trajectory of a moving object in a image sequence. Instead, it performs tracking by combining two powerful stochastic modeling techniques: The first one is the technique of so-called Pseudo-2D Hidden Markov Models (P2DHMMs) used for capturing the shape of an object within an image frame, and the second technique is the well-known Kalman-filtering algorithm, that uses the output of the P2DHMM for tracking the object by estimation of a bounding box trajectory indicating the location of the object within the entire video sequence. Both algorithms are cooperating together in an optimal way, and with this cooperative feedback, the proposed approach even makes the tracking of objects possible in the presence of background motions, for instance caused by other moving objects such as cars, or by camera operations as e.g. panning or zooming. Thus, the test sequence provided for PETS2000 represents a suitable testbed for our system, because it contains different moving objects and thus confrontates the system with the challenge to track a specifically selected object in presence of other objects that are moving across the entire image and are even intersecting with the trajectories of other objects in the sequence. Our results confirm that our approach can handle the difficult task of tracking a specific person within the image sequence without losing its trace despite the previously mentioned problems caused by the other moving object in the sequence. The advantages of our tracking approach would have been even more apparent if the test sequence was acquired with a moving camera, causing motion information distributed all over the entire image sequence. Of course, our system has also certain limitations. It is not (yet) designed as a general "alert-system" that can detect different moving objects in a real-world scenario and track them simultaneously. Instead, it has to be initialized on one specific object, but is then capable of tracking this one reliably despite disturbing motions in the image sequence. In the paper, we will report on the advantages and limitations of our statistical object tracker.